On May 19, 2026, Andrej Karpathy posted a 54-word tweet on X: he was joining the Anthropic Pre-training team. No long essay, no preamble—within 40 minutes, the Hacker News post had 230+ upvotes. His mission is to build a new team that uses Claude itself to accelerate Claude's pre-training research—a public bet that "AI-assisted research, not pure compute scaling, is the decisive factor in the next phase of competition."
Andrej Karpathy's resume is practically a capsule history of modern deep learning. In 2015, he joined OpenAI as a founding member, before the name GPT even existed; in 2017, he was recruited by Elon Musk to Tesla, leading the Autopilot and FSD projects, completing his transition from researcher to engineering practitioner; in 2022, he briefly returned to OpenAI, right during the most intensive infrastructure build-out just before the ChatGPT explosion, but stayed only about a year, leaving again in 2024 to found Eureka Labs, focused on reinventing education with AI assistants. Outside of companies, he is one of the world's most influential AI science communicators—his YouTube course Neural Networks: Zero to Hero has enabled countless self-learners to truly understand the inner workings of the Transformer.
According to TechCrunch, Karpathy joined the Pre-training team, reporting directly to team lead Nick Joseph. His core mission: build a new team dedicated to researching how to use Claude itself to accelerate Pre-training research. Pre-training is the stage where Claude acquires language capabilities and world knowledge from scratch; it is the most expensive, compute-intensive phase of the entire training pipeline, and the one with the deepest impact on final model quality—breakthroughs in this stage directly determine the ceiling for the next generation of models. What Karpathy is doing is using Claude as a research tool, letting AI help optimize AI training itself—precisely one of the most aggressive research directions in current AI companies ("AI for AI research"). Axios's analysis pinpointed the strategic logic: "Karpathy is one of the few AI figures who holds credibility simultaneously in research, industry, and education. Recruiting him to build such a team makes it clear that Anthropic believes AI-assisted research, not pure compute, is how it will compete with OpenAI and Google."
This hire occurred against the backdrop of Anthropic's rapid rise as a top AI talent magnet. On the very same day Karpathy announced his arrival, veteran cybersecurity expert Chris Rohlf (who previously led core security work at Yahoo and Meta, with over 20 years of industry experience) joined Anthropic's Frontier Red Team, specifically stress-testing high-risk AI systems. Two people, two directions—one pointing toward the ceiling of model capability, the other toward the floor of model safety. Anthropic is pushing hard on both ends simultaneously. Looking more broadly, the very top researchers are no longer circulating only between OpenAI and DeepMind; Anthropic has become another genuine destination.
He offered a key sentence in his tweet: "I think the next few years at the frontier of LLMs will be especially formative."—years that are especially formative. This is a judgment about timing, not simply an endorsement of a particular company. His choice of Anthropic over returning to OpenAI is also telling: when he left OpenAI in 2024, there was already considerable external discussion about its internal culture and governance issues; Anthropic, with "AI safety" as its core narrative, aligns more closely with his consistently cautious, pro-interpretability stance. He also mentioned that his passion for education won't disappear, "planning to return to this work at the appropriate time"—the Eureka Labs story isn't over, just temporarily on hold.
Over the past year or so, while not on any big company's payroll, he maintained a remarkably high density of intellectual output: in early 2024, he popularized the "vibe coding" concept, using natural language to describe intent and letting AI generate code; in a March 2026 podcast, he admitted he hadn't handwritten a single line of code since December 2025, with the AI handling proportion climbing from 20% to 100%, calling this state "AI psychosis" and predicting 2026 would be the "year of slopacolypse." These insights are the intellectual assets he brings to the Anthropic Pre-training team—a researcher who treats himself as an extreme user of AI toolchains, building a team focused on "using Claude to accelerate AI research," represents a dual alignment of direction and personality.
Only by placing this move within a larger trend can we grasp the true weight of this jump.
"Recursive self-improvement" was once a warning from safety researchers; today it is an engineering task that Anthropic openly acknowledges pursuing—Karpathy's mandate is precisely the engineering implementation of RSI's first phase.
Once pre-training research is accelerated by AI, a positive feedback loop kicks in: better models → faster research → better next-generation models. Only a handful of labs with compute, data, and top researchers can enter this flywheel.
The pricing power of mid-level researchers is declining, while pricing power at the very top is rising—those willing to forgo management roles to be individual contributors show that the opportunity premium of frontier labs has grown large enough to compensate for everything.
The OpenAI vs. Anthropic competition has spread to the level of research paradigms: rapid iteration and broad deployment vs. safety research alongside capability development. Other players will find it increasingly difficult to find a foothold between the two poles.
A sharper voice came from X commenter roon (209K views): "Ironically, this year might be a sad one for AI researchers—they are the first to be affected on the RSI hot path, and the market's pricing power for them will likely shrink as this generation of models commoditizes their skills." Karpathy joining Anthropic to "use Claude to accelerate Pre-training research" is, in a sense, personally accelerating this very process.
First, he chose Pre-training—not Fine-tuning, RLHF, or the application layer, betting that breakthroughs in foundational capabilities have not yet hit a ceiling. Second, he chose using Claude for AI research—betting that AI-assisted research will become the next core competitive dimension, not merely compute scaling. Third, he chose Anthropic over returning to OpenAI—betting that the route balancing safety and capability holds an advantage in long-term competition. All three bets are judgments about where AI is heading in the coming years. Time will tell whether his intuition is right once again.
First published 2026-07-24